Actuarial Notes

Insurer Operational Data

Study guide: Exam DISC-DA

Insurer Operational Data is the internal data an insurer's own systems record as it writes and services policies and settles claims: a policy database of exposures, premium and risk characteristics, a claims database of every payment and reserve change, and accounting records of expenses that belong to no single policy. It is the main source for pricing and modelling an existing product, because it describes the insurer's own book at the level of the individual policy and claim.

\text{WP}_p = \sum_{r \in p} \text{WP}_r

\text{Reported}_c(t) = \sum_{\tau \le t} \text{Paid}_{c,\tau} + \text{Case}_c(t)

  • Symbols: The first block nets a policy p from its records r: written premium \text{WP}_r (and, the same way, written exposure) is recorded once at issue and again, positive or negative, for every cancellation or mid-term change, so the policy's figure is the sum of its records. In the second block, the reported (case-incurred) loss on claim c at valuation date t is the sum of its payments \text{Paid}_{c,\tau} made on or before t plus the case reserve \text{Case}_c(t) still held on it at t.
  • Two kinds of internal data: Risk information (exposures, premium, claim counts, losses and the characteristics of each policy and claim) is held record by record. Accounting information, such as underwriting expenses and unallocated loss adjustment expenses, is generally available only in aggregate, tracked by calendar year and sometimes approximated by line and state.
  • Policy database: Each record is a policy or a subdivision of one, and its fields describe it. What counts as a record depends on the exposure measure and on how premium is calculated: in homeowners a record may be one home for an annual term, in U.S. workers compensation records are often kept by payroll classification, and in personal auto a record may be kept for each coverage, each car and each driver. Typical fields are the policy identifier, risk identifiers (such as vehicle and operator numbers), the original effective and termination dates and the transaction date, written premium and written exposure by coverage, and the characteristics: rating variables, underwriting variables and anything else known about the risk.
  • Transactions, not snapshots: A cancellation or a mid-term change does not overwrite the original record. It adds records: one that reverses the unearned part of the original terms and, for a change, one that writes the rest of the term on the new terms. Earned and in-force premium are then calculated from the records.
  • Claims database: Each record is usually a transaction on one claim, such as a payment or a change in its reserve. Typical fields are the policy and risk identifiers (to match the claim to its policy record), the claim and claimant identifiers, the accident, report and transaction dates, the claim status (open, closed and, where common, reopened), the claim count, the paid loss, an event identifier for catastrophes, the case reserve, paid allocated loss adjustment expense, salvage and subrogation recoveries, and claim characteristics such as the type of injury.
  • Bringing them together: Because policy and claim information sit in separate databases, an analysis matches them on the policy and risk identifiers with a table join, and aggregates the transactions by calendar year, accident year, policy year or report year. Some insurers provide a data mart built for ratemaking; elsewhere the analyst starts from the transactional databases and reshapes the data.
  • Strengths: The data are the insurer's own experience, at the finest level of detail kept, so they can be summarized for an overall rate review or kept at the policy level for a multivariate classification analysis.
  • Limits: A characteristic is often captured in several systems and can be hard to merge. A field captured as a changing value (a driver's age) is less useful than a stable one it can be derived from (the date of birth). A claim characteristic can describe what drives claim size, but only characteristics known for every policyholder when the policy is quoted can be used in a rating algorithm. A new product has no history of its own, and sparse or unstable internal data may need external data or industry statistical data as a benchmark. The data must still be reviewed for quality before use.

Building Policy Records for a Cancellation and a Mid-Term Change A homeowners insurer writes two annual policies, earned evenly by month.

  • Policy X is written on March 1, 2024 for $900 and cancelled on December 1, 2024.
  • Policy Y is written on July 1, 2024 for $800 with a $500 deductible. On January 1, 2025 the insured raises the deductible to $1,000, and the full annual premium on the new terms is $700.

List the transaction records each policy produces, with written exposure and written premium, and give each policy's net figures.

Answer Policy X ran 9 of its 12 months, so the cancellation record reverses the last quarter of the term.

\begin{align*}
\text{WP}_X &= 900 - 0.25 \times 900 \\
&= 900 - 225 \\
&= 675
\end{align*}

X has 2 records: +1.00 exposure and +\$900 on March 1, 2024, then -0.25 and -\$225 on December 1, 2024. Net, it is 0.75 exposure and $675.

Policy Y is changed halfway through its term. One record reverses the unearned half on the old deductible, and another writes that half on the new deductible.

\begin{align*}
\text{WP}_Y &= 800 - 0.50 \times 800 + 0.50 \times 700 \\
&= 800 - 400 + 350 \\
&= 750
\end{align*}

Y has 3 records: +1.00 and +\$800 at a $500 deductible on July 1, 2024; -0.50 and -\$400 at $500 on January 1, 2025; and +0.50 and +\$350 at $1,000 on January 1, 2025. Net, it is 1.00 exposure and $750. For a model of the deductible's effect, the records are collapsed by period of exposure: 0.50 exposure and $400 at the $500 deductible, and 0.50 and $350 at $1,000, so each piece carries the characteristics in force while it was earned.

Reading a Claim's Transaction Records A loss on Policy Y occurs on November 20, 2024 and is reported on November 25, 2024, when a $12,000 case reserve is set. On January 15, 2025 the insurer pays $3,000 and lowers the reserve to $9,000. On April 10, 2025 it pays $8,000 and closes the claim. On June 1, 2025 it recovers $2,500 by subrogation. How many records does the claims database hold for this claim, what is the reported loss at December 31, 2024 and at December 31, 2025, and what does the claim contribute to calendar year 2025 reported losses?

Answer Each transaction is a record: the report with its reserve, the two payments with their reserve changes, and the recovery, so the claim has 4 records under one claim identifier.

\begin{align*}
\text{Reported}(\text{12/31/2024}) &= 0 + 12{,}000 \\
&= 12{,}000 \\
\text{Reported}(\text{12/31/2025}) &= 3{,}000 + 8{,}000 + 0 \\
&= 11{,}000 \\
\text{CY 2025 reported} &= 11{,}000 + (0 - 12{,}000) \\
&= -1{,}000
\end{align*}

Accident year 2024 shows $11,000 of reported loss at the end of 2025, gross of the $2,500 recovery, which is tracked separately and linked to the claim. Calendar year 2025 sees payments of $11,000 but releases the $12,000 reserve set in 2024, so this claim lowers calendar year 2025 reported losses by $1,000. The same transactions give different totals depending on how they are aggregated.

How it appears on the exam

  • DISC-DA names insurer operational data as the first section of its topic on collecting data for insurer models, beside statistical plans and external data sources. Application questions describe a modelling need and ask which internal source holds the data (the policy database, the claims database or accounting records), what one record represents, or why a field cannot be used as intended.
  • Expect scenarios that turn on the transaction structure: a cancelled or amended policy, a claim with several payments, or a claim characteristic proposed as a rating variable. Topic 9 of the course builds claims and underwriting modelling data sets from these same sources.

Exam traps

  • A record is not a policy: One policy can have many records (per coverage, car, driver, or mid-term change), and one claim many transaction records. Counting rows counts records; count distinct policy or claim identifiers instead.
  • A cancellation adds a record: The original record keeps its full-term exposure and premium, and the cancellation is a negative record. Summing only the issue records overstates both.
  • Known at quotation: A characteristic recorded on the claim, such as the type of injury, may explain severity in a reserving study, but it cannot be a rating variable, because it is not known for a policyholder before a claim.
  • Expenses are not on the records: Underwriting expenses and unallocated loss adjustment expenses are tracked in aggregate, not by policy or claim. Allocated loss adjustment expense is the part of claim expense that is on the claim record.